Because the problem was never the prompt. It is the file your AI reads before the prompt, and most business owners have never built one. This article answers why AI output still sounds like nobody wrote it, and what to build instead.
The week four unrelated groups said the same thing
Last week a woman in my community told me she had spent a year collecting prompts.
Hundreds of them. Organized in folders. Color coded. She had done everything the internet told her to do, and her AI still wrote like a press release from a company that does not exist.
She asked me what prompt she was missing.
Here is the honest answer, and it is going to sting a little: you are not missing a prompt. You are missing a memory. Generic output is a context problem, not a prompting problem. The fix is not a better instruction; it is a small set of persistent files that your AI reads before it reads anything you type.
I did not arrive at that on my own. Molly Mahoney published a guide on September 2, 2026 called "Stop Training Your AI From Scratch Every Day", and she named the exact five documents: brand voice, ideal client, offers, company details, and story bank. That framing is hers, and it is the cleanest version of this idea I have seen written for people who do not code.
What made me sit up was the timing.
In the same seven days, a developer tool launched on Product Hunt built to turn coding sessions into persistent rules and skills. A subreddit full of daily Claude users converged on the same procedural advice. And a research paper hit arXiv describing an agent runtime that throws away the growing transcript and keeps a lean, mutable state file instead.
A marketing educator. A software company. A forum. A research team. Four vocabularies, one architecture.
The business owner version is not technical. It is five documents in a folder.
Key Takeaways
- Generic AI output is a memory and context problem, not a prompt-quality problem.
- Stanford and Berkeley researchers documented that language models retrieve information best at the beginning and end of a context window and get measurably worse in the middle.
- In Brafton's 2026 survey of 132 marketers using AI, "thin or generic-sounding" was the single most cited content quality concern.
- Software teams solved this with version-controlled context files that the agent reads first; only 5 percent of mature open-source repositories have adopted one.
- Molly Mahoney's five files, brand voice, ideal client, offers, company details, and story bank, are the non-technical version of the same architecture.
The problem: you have been optimizing the wrong layer
Prompt engineering was the right advice in 2023.
It stopped being the whole answer somewhere around the time these tools got long memories and started forgetting things inside them.
Here is the thing most business owners never get told. When you open a fresh chat, your AI knows nothing about you. Not your voice. Not who you serve. Not what you sell. Not the story you tell on stage about the worst quarter of your life. So it reaches for the average of everything it has ever read, and the average of everything is, by definition, generic.
You compensate by writing a longer prompt. That works, sort of, for one output.
Then you open a new chat tomorrow and do it all again. And again. You are re-teaching the same business facts every single morning, paying for it in time, and getting a slightly different result each time because you never phrase it identically twice.
That is not a skill gap. That is a system design flaw.
I know something about rebuilding from a blank slate. I have started over from nothing more than once, including from a place most people do not put in a bio, and the lesson that stuck was this: if you have to reconstruct the same foundation every morning, you will never build anything on top of it. You are not lazy. You are just spending all your energy on the first floor.
The second problem is worse, and it is the one almost nobody sees coming.
The advice you get is "just paste more context in." So you dump a fifty-page brand document into the chat window and assume the model absorbed it. It did not. Not reliably. Not the parts in the middle.
That is not an opinion. That is peer-reviewed.
The evidence: what the research actually says
Models do not read your context evenly. In "Lost in the Middle: How Language Models Use Long Contexts," Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang, working across Stanford, UC Berkeley, and Samaya AI, tested how well models find relevant information depending on where it sits in the input. Their finding, published in Transactions of the Association for Computational Linguistics, is that performance is highest when the relevant information sits at the beginning or the end of the context, and degrades significantly when the model has to reach into the middle. That held even for models explicitly built for long contexts.
Translation for the rest of us: page 27 of your brand guide might as well be blank.
It gets less reliable the more you paste. Chroma's context rot research evaluated 18 language models, including GPT-4.1, Claude 4, Gemini 2.5, and Qwen3. Their conclusion was blunt: models do not use their context uniformly, and performance grows increasingly unreliable as input length grows, even on simple tasks. More input is not more understanding.
And marketers feel it in the output. Brafton surveyed 132 marketers currently using AI in their workflows and asked what worried them most about content quality. The top answer, chosen by 87 of them, was "the content is thin or generic-sounding." It beat factual accuracy. It beat SEO. Forty-three said it does not reflect their expertise. The people most likely to flag it were marketers with eleven or more years of experience, the ones who can tell the difference between polished and differentiated.
Software teams already built the fix. Developers hit this wall a year earlier and responded by inventing a file format. Seyedmoein Mohsenimofidi, Matthias Galster, Christoph Treude, and Sebastian Baltes studied its adoption in "Context Engineering for AI Agents in Open-Source Software," presented at the 2026 Mining Software Repositories conference. They scanned 10,000 mature GitHub repositories and found that only 466, roughly 5 percent, had adopted an AI context file at all. The files that exist average 142 to 310 lines depending on format. Half of them have never been edited since the day they were created.
Read that again. Five percent adoption among professional software teams. This is early for everybody.
The research lab version. On August 26, 2026, Sanket Badhe, Priyanka Tiwari, and Jonghyun Chung published "SKILL.state: Scalable Long-Horizon Agent Skills", accepted at EMNLP. Their argument is that agent runtimes fail on long jobs because they keep appending everything to an ever-growing conversation history, which causes latency degradation and what they call context-poisoning failures. Their fix is to replace the transcript with an explicit, mutable state file. At each step the model sees only the fixed skill specification, the current state, and the latest observation. Everything else gets discarded.
The tool version. On September 3, 2026, Blume.codes launched on Product Hunt. Co-founder Peder Aaby wrote in the launch thread that they pivoted their entire company after their previous startup suffered from agent drift, and that even with proper rules and docs in place, "the setup rotted fast" as the repo grew. Their product watches sessions and turns repeated corrections into persistent rules.
Same disease. Same prescription. Four vocabularies.
The solution: five files, read before the prompt
Molly Mahoney's list is the one I hand to business owners, because it is written in the language of a business and not the language of a repository.
Brand voice. Not adjectives. Actual rules. Sentence length. Words you never use. Words you always use. Three paragraphs of your own writing pasted in raw as reference samples. "Professional yet approachable" is worthless. "Never use the word leverage as a verb" is a rule a machine can follow.
Ideal client. Who they are, what keeps them up, what they have already tried, what they are skeptical of, and the exact phrases they use to describe their own problem. That last one matters more than the demographics.
Offers. Every product, what it costs, who it is for, who it is not for, and what result it actually produces. Include what you do not sell. Boundaries are context too.
Company details. Founding story, positioning, the promise, the guarantee, the things you will not do. This is the file that stops your AI from inventing services you do not offer.
Story bank. This is the one people skip and it is the one that does the most work. Your customer wins. Your failures. The line you use on stage. The moment the business almost died. Generic writing is writing with no specifics in it, and specifics live in stories. An AI cannot make yours up. It can only retrieve them if you wrote them down.
Now the part that separates this from a folder of documents nobody opens.
These files have to be loaded before the request, not pasted into the middle of a long chat. Put them in the project instructions, the custom instructions, or a project knowledge base, whichever your tool calls it. That is the whole point of what Liu and his coauthors found. Position matters. Front-load the foundation, then ask.
And keep them short. The software researchers found working context files running 142 to 310 lines. Not fifty pages. If your brand voice file is longer than two pages, you have written a brochure, not a specification.
Here is what changed for me. I stopped opening a blank chat. I open a workspace that already knows who I am, and my first message is a request instead of a briefing. The output difference is not subtle. It is the difference between an assistant and a stranger.
How to build it this week
1. Stop and audit one output. Take the last thing your AI wrote that made you cringe. Highlight every sentence that could have been written about any business in your category. That highlighted text is your missing context, and it will point you straight at which file to write first.
2. Write the story bank first. Everyone starts with brand voice. Start here instead. Open a document and write fifteen things that actually happened: a client result, a bad launch, a conversation that changed your pricing, the thing your mentor told you. Two to four sentences each. This is the file that makes output sound like a person.
3. Build the other four with an interview, not a blank page. Do not stare at an empty document. Use this prompt, and answer the questions it asks you out loud:
[The Job]
Interview me and then write my brand voice context file.
[The Background]
I run a small business and I want a permanent reference document my AI
reads before every writing task. I have pasted three samples of my own
writing below. The file needs to capture how I actually sound: sentence
rhythm, words I use, words I avoid, punctuation habits, and how I open
and close a piece.
[The Deliverable]
A brand voice file under 500 words, written as specific rules a machine
can follow, not adjectives. Include a "never do this" section.
[The Questions]
Ask me any questions you have.
4. Load them where the tool reads first. Project instructions in Claude. Custom instructions or a project in ChatGPT. A Gem in Gemini. Whatever your platform calls the persistent layer, that is where these go. Pasting them into a chat and hoping is the failure mode.
5. Clear between unrelated tasks. Do not run your Q4 planning session and your email newsletter in the same thread. The Chroma research is clear that longer inputs get less reliable. Fresh thread, persistent files, new request.
6. Update the files when you correct the AI. This is the whole insight behind Blume.codes. When you find yourself giving the same correction twice, that correction belongs in a file, not in a chat. Half the context files in that GitHub study were never edited after creation. Do not build a museum.
7. Reread them once a quarter. Your offers change. Your positioning sharpens. Your stories accumulate. A stale context file quietly makes your AI wrong instead of just generic, which is worse.
Frequently Asked Questions
Is this the same thing as a custom GPT or a Claude Project?
Close, but not identical. A custom GPT or Project is the container. The five files are what you put inside it. Most people build the container and leave it nearly empty, then wonder why the output matches the default. The container without the context files gives you nothing new.
How long should each file be?
One to two pages. Researchers studying real AI context files in open-source projects found them averaging between 142 and 310 lines. Longer is not better, because models retrieve less reliably from long inputs. Write specifications, not brochures, and cut anything that is not an instruction or a fact.
Do I still need to write good prompts?
Yes, but the job changes. With context files loaded, your prompt only has to describe the task, not re-teach your business. Prompts get shorter and results get more consistent. You are moving effort from repetition into a one-time build.
Will this stop AI content from sounding like AI?
It removes the biggest cause. In Brafton's 2026 survey, 87 of 132 marketers named thin or generic output as their top concern, and generic means "contains no specifics." Your story bank and client file supply the specifics. You still edit before you publish.
What if I use several different AI tools?
Write the files once in plain text or markdown and paste them into each platform's persistent instruction area. Developers standardized on a shared file format for the same reason. Keep one master copy so you update in one place instead of five.
The close
She had hundreds of prompts and no memory.
That is most business owners right now, and it is not a character flaw. It is what happens when the entire internet teaches you to optimize the request and nobody teaches you to build the foundation the request lands on.
Four completely unrelated groups figured this out in the same week, in four different vocabularies, because it is the natural next problem once these tools get good enough to use daily. A marketing educator called it training your AI once instead of every day. A developer tool called it rules and skills. A subreddit called it a project context file. A research team called it mutable execution state.
They are all describing the same thing. Give the machine a memory before you give it a job.
Candidly, this is not the exciting version of AI advice. Nobody builds a following teaching people to write a two-page document about their own offers. But it is the difference between a tool that sounds like you and a tool that sounds like the internet's average opinion of a business like yours.
Give me credit where it belongs: Molly Mahoney named these five files in plain English before I did, and she named them for people who do not write code. That is the version that actually helps.
Write the story bank first. Do it this week.
Your AI cannot tell your stories if you never wrote them down.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he teaches non-technical entrepreneurs to use AI to amplify the skill and experience they already have. He runs the AI Insiders membership and live AI trainings, and speaks on practical AI for business owners. He rebuilt his own business from a blank slate more than once, which is probably why the idea of starting from scratch every single morning bothers him as much as it does.
Sources
- Stop Training Your AI From Scratch Every Day, Molly Mahoney, September 2, 2026
- Lost in the Middle: How Language Models Use Long Contexts, Liu et al., TACL 2024
- Context Rot: How Increasing Input Tokens Impacts LLM Performance, Chroma
- The 4 Biggest Challenges in AI Content Creation, Brafton 2026 AI marketing survey
- Context Engineering for AI Agents in Open-Source Software, Mohsenimofidi et al., MSR 2026
- SKILL.state: Scalable Long-Horizon Agent Skills, Badhe, Tiwari, and Chung, arXiv 2608.26263
- Blume.codes on Product Hunt, September 3, 2026
